Deep Learning based adaptive characterisation of QPUs

Detailed device characterisation is necessary for both obtaining the optimised gate-sets on a given hardware as well as identifying device imperfections and error sources to improve the next design iteration. Typical textbook characterisation routines do not scale efficiently to large multi-qubit chips, requiring the development of techniques based on statistical and information theoretic foundations. We present the application of Deep Learning based characterisation techniques that adaptively recommends Bayesian Optimal Experiments at every step to maximise the expected information gain about the system, while taking into account the entire history of past experiments. The cost of calculating expensive Bayesian posteriors is amortised by the use of a Reinforcement Learning system which simultaneously learns both the design policy and lower bounds on the otherwise computationally intractable Expected Information Gain. A physics accurate fully-differentiable digital twin that models the open quantum dynamics of the QPU, the control electronics and the noise & transfer functions for the whole stack lies at the heart of this closed loop adaptive calibration and characterisation process. We demonstrate the application of these bayesian adaptive experiments on multi-qubit superconducting QPU systems.